The market for uncomfortable truths

Executives have started asking AI for the candor they no longer expect from their advisers. And for that, they have chosen the only interlocutor trained to tell them they are right. How did we reach a point where the machine seems more honest than the person we pay to be honest with us?

The shift is measurable. A BCG report notes that some CEOs already use AI to pressure-test decisions and flag risks their human advisers may be “too polite, too biased, or too removed from the problem” to raise ¹. And by the end of 2024, 44% of executives at large U.S. companies said they would change a decision they had already planned to make based on what AI told them ².

But the remedy reverses the premise. Models are fine-tuned to produce the answers people prefer —reinforcement learning from human preferences—. And we prefer to be told we are right. Anthropic documented it: agreement with the user’s views is one of the strongest predictors of which answer we will prefer, and its own preference model chooses the sycophantic answer over the truthful one 95% of the time ³. And it gets worse as the model gets larger ⁴. It is not a bug. It is the objective working as designed.

There is a further problem: the harder the case, the worse we are at distinguishing truth from flattery ³. And hard cases are precisely why people hire advisers. In April 2025, OpenAI rolled back an update to GPT-4o in four days: it was not failing; people liked it too much. The company had optimized the thumbs-up signal and produced a model that was “overly supportive but disingenuous” ⁵. This is not an accusation from AI critics. It comes from the people building it.

AI knows how to frame the objection. What it cannot do is pay a price for raising it.

Because the risk does not accompany the message. It is the message. Anyone who contradicts the person paying them puts the renewal, the relationship, and their reputation at stake. That is the signal: someone weighed the cost and spoke anyway. It would be dishonest to stop there. We, too, have been optimized for stated client satisfaction: for twenty years, we have adjusted our advice to fit client preferences and called it client orientation. I wrote recently that judgment is defended by making the decision process visible; that was about what you know, while this is about what you are willing to say. And disagreement is not declared. It is demonstrated —recommendations not to act, objections put in writing, assignments turned down— and only someone willing to lose accounts can demonstrate it.

The client who flees an overly polite adviser for a machine trained to please them has not misdiagnosed the problem. They have chosen the wrong remedy. Diagnosis has been commoditized. The uncomfortable truth has not: it still costs the person who speaks it something, and that cost is precisely what the client is buying. Look back at the last proposal you won: what would it have cost you to say what you left unsaid?